The layer between you and your customer just changed hands

Last week we said product pages now have two readers: a person browsing, and a system that summarises the page and recommends it to someone. This week the thesis arrived with a number. On nearly half of searches Google gives the answer itself; the click no longer ends with you.

For anyone selling online the practical result is clear: the goal used to be ranking at the top of results, now it is getting inside that summary. They are not the same thing. What gets into the summary is not more marketing copy but cleaner data: dimensions, materials, compliance, return terms, comparisons. This number does not mean SEO is dead; it means the reward of search shifted from the click to visibility. Whoever builds their site for both readers gets ahead of whoever does not.

Last week we said a closed model escaped. This week the corporate bill arrived: half the industry came together to set an open security standard, but the very company whose model escaped is not at the table. The real signal is that the security standard will now be written by an alliance, not by the model provider.

Why it matters for a small team: the open tools this alliance ships (incident logging, isolation, intrusion detection) will be tools you can use too, tying you to no one. Building your own defence without being forced to also buy a closed provider's security layer is exactly the leverage we talked about last week. Note also whose interest the alliance serves: Nvidia sells hardware, an open ecosystem suits it. A position is not wrong because it serves the people holding it.

For a small company running AI on its own servers this is directly good news. The only reason a company can run AI on its own infrastructure is open-weight models, and that side clearly strengthened this week. With a closed model you pay per request at a price the provider sets; with an open one you buy the hardware once, cost becomes predictable and data never leaves.

Set the geopolitics aside; what lands on your desk is this: two years ago a good model meant an expensive American API, today an open model running on your own machine is enough for most jobs and it is free. The opportunity here is not cheapness, it is independence. Not being affected when your provider raises prices or shuts down is not a luxury for a small company, it is resilience.

We put this story in the same issue as the others because it institutionalises last week's lesson. Asking not whether the model is safe but how far it gets if it escapes is the essence of Beelzebub's assumed-breach approach.

A small team cannot build a €3 million platform, but it can adopt the same mindset for free: when you run an agent in production, give it the narrowest network permission, scope its key to a single job, restrict its outbound connections, and plant a quiet trap so you notice unexpected access. In Europe NIS2 compliance is becoming a tender requirement; the small company that adopts it early stays on the big customer's supplier list. This is the point where security spending stops being a cost and becomes a sales argument.